pkg: drop hsa-rocr dep, warn at install time instead (#300)

The relocatable DEB/RPM hard-depended on hsa-rocr, but ROCm 7.13 no
longer ships a package by that name -- install fails on TheRock 7.13
with "amdrocm7-transferbench depends on hsa-rocr; however: Package
hsa-rocr is not configured yet" (ROCM-24669).

The dep also never made sense for the relocatable artifact: it is
built from the TheRock SDK and meant to install on systems where ROCm
came from a tarball and is not tracked by apt/dpkg at all. Any apt
dep on a ROCm component breaks that audience.

Drop hsa-rocr from CPACK_DEBIAN_PACKAGE_DEPENDS,
CPACK_RPM_PACKAGE_REQUIRES, and the legacy
rocm_package_add_dependencies call. Wire a shared scriptlet
(packaging/postinst-check-hsa.sh) as the DEB postinst and the RPM
%post that probes ldconfig + the standard /opt/rocm{,-*,/extras-*,
/core-*}/lib prefixes for libhsa-runtime64.so.1 and prints a stderr
warning when none of them have it. Always exits 0 so install never
fails on its account -- the warning surfaces the missing runtime at
install time instead of as a dynamic-linker error on first run.

numactl / libnuma1 stay in the dep list since those are real OS
packages that exist independent of any ROCm install.

Author: thananon

CMakeLists.txt

@@ -360,7 +360,7 @@ if(BUILD_RELOCATABLE_PACKAGE)
   # DEB
   set(CPACK_DEBIAN_PACKAGE_NAME       "${CPACK_PACKAGE_NAME}")
   set(CPACK_DEBIAN_PACKAGE_ARCHITECTURE "amd64")
-  set(CPACK_DEBIAN_PACKAGE_DEPENDS    "numactl, libnuma1, hsa-rocr")
+  set(CPACK_DEBIAN_PACKAGE_DEPENDS    "numactl, libnuma1")
   set(CPACK_DEBIAN_PACKAGE_MAINTAINER "${CPACK_PACKAGE_CONTACT}")
   if(NOT "${_tb_pkg_release}" STREQUAL "")
     set(CPACK_DEBIAN_PACKAGE_RELEASE "${_tb_pkg_release}")
@@ -373,7 +373,7 @@ if(BUILD_RELOCATABLE_PACKAGE)
   # RPM
   set(CPACK_RPM_PACKAGE_NAME    "${CPACK_PACKAGE_NAME}")
   set(CPACK_RPM_PACKAGE_LICENSE "MIT")
-  set(CPACK_RPM_PACKAGE_REQUIRES "numactl, hsa-rocr")
+  set(CPACK_RPM_PACKAGE_REQUIRES "numactl")
   set(CPACK_RPM_PACKAGE_VENDOR  "${CPACK_PACKAGE_VENDOR}")
   if(NOT "${_tb_pkg_release}" STREQUAL "")
     set(CPACK_RPM_PACKAGE_RELEASE "${_tb_pkg_release}")
@@ -394,6 +394,17 @@ if(BUILD_RELOCATABLE_PACKAGE)
         "${_rpm_exclude_prefix}"
         "${_rpm_exclude_prefix}/bin")
 
+  # Advisory install-time check for libhsa-runtime64.so.1. The package declares
+  # no hard ROCm dep so it can install on TheRock-tarball systems where no ROCm
+  # component is tracked by apt/dpkg; the postinst warns (never fails) when the
+  # HSA runtime is not discoverable, so a missing runtime surfaces at install
+  # time instead of as a dynamic-linker error on first invocation.
+  set(_tb_postinst_src "${CMAKE_CURRENT_SOURCE_DIR}/packaging/postinst-check-hsa.sh")
+  set(_tb_postinst_deb "${CMAKE_BINARY_DIR}/packaging/postinst")
+  configure_file("${_tb_postinst_src}" "${_tb_postinst_deb}" COPYONLY)
+  set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${_tb_postinst_deb}")
+  set(CPACK_RPM_POST_INSTALL_SCRIPT_FILE "${_tb_postinst_src}")
+
   # TGZ — embed release tag so successive runs do not collide on the same key.
   # CMake 3.13+ honors CPACK_ARCHIVE_FILE_NAME for archive generators, but
   # CMake 3.22 (Ubuntu 22.04) falls back to CPACK_PACKAGE_FILE_NAME for TGZ.
@@ -414,7 +425,7 @@ else()
   rocm_setup_version(VERSION ${VERSION_STRING})
 
   # Package specific CPACK vars
-  rocm_package_add_dependencies(DEPENDS "numactl" "hsa-rocr")
+  rocm_package_add_dependencies(DEPENDS "numactl")
 
   set(CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/LICENSE.md")
   set(CPACK_RPM_PACKAGE_LICENSE "MIT")

packaging/postinst-check-hsa.sh

@@ -0,0 +1,50 @@
+#!/bin/sh
+# Advisory check used as both the DEB postinst and the RPM %post scriptlet.
+# TransferBench links against libhsa-runtime64.so.1; without it the binary
+# fails at first run with a dynamic-linker error. The relocatable package
+# declares no hard dep on hsa-rocr (or any ROCm component) because it is
+# expected to install on TheRock-tarball systems where no ROCm package is
+# tracked by apt/dpkg. Warn here so the user diagnoses a missing runtime
+# at install time, not at TransferBench launch.
+#
+# Always exits 0 — this is advisory, never fatal.
+
+set -e
+
+found=0
+
+if command -v ldconfig >/dev/null 2>&1; then
+    if ldconfig -p 2>/dev/null | grep -q 'libhsa-runtime64\.so\.1'; then
+        found=1
+    fi
+fi
+
+if [ "$found" -eq 0 ]; then
+    for d in /opt/rocm/lib /opt/rocm/lib64 /opt/rocm-*/lib /opt/rocm/extras-*/lib /opt/rocm/core-*/lib; do
+        if [ -e "$d/libhsa-runtime64.so.1" ]; then
+            found=1
+            break
+        fi
+    done
+fi
+
+if [ "$found" -eq 0 ]; then
+    cat >&2 <<'EOF'
+====================================================================
+TransferBench: WARNING
+
+libhsa-runtime64.so.1 was not found on the dynamic loader path or
+under any of /opt/rocm/lib, /opt/rocm-*/lib, or /opt/rocm/extras-*/lib.
+
+TransferBench requires the ROCm HSA runtime at run time. Install a
+ROCm 7.x stack (system packages or TheRock SDK) before invoking
+TransferBench, or set LD_LIBRARY_PATH to a directory containing
+libhsa-runtime64.so.1.
+
+Without it, TransferBench will fail at startup with:
+    error while loading shared libraries: libhsa-runtime64.so.1
+====================================================================
+EOF
+fi
+
+exit 0